I use OpenSCAD with Claude Code and it's pretty impressive. In practice the flow of letting the machine just spend its time iterating on the design with PNGs that it rotates and renders before it actually gives me an STL works quite well.
These are fun but as long as it’s OpenSCAD it will always be a hobbyist toy and not able to be used for professional use. Which is fine! But imo an openscad rendering pipeline is pretty easy. I’m more interested in someone trying to take a stab at vibe coding models that output Brep compatible formats and generate STEP files. This is much, much harder to do but is the main step towards doing this for actual professional use cases.
Presumably someone is getting closer to this, curious who the most robust player in that space is.
Also curious if building an actual kernel replacement for open cascade is on the table now with AI, it’s a very tough thing to do but now it seems somewhat tractable in 2026
I have been trying to force LLMs to work with geometries for over a month and it's so hard. Even the best LLMs have an extremely poor sense of geometric relationships in my testing. I would also stay away from mesh based CAD like OpenSCAD and go straight for build123d which operates on real solid models (BREP): https://build123d.readthedocs.io/en/latest/index.html
I also been trying to use LLM for creating house plans but it got bad sense of directions and spaces and sizes and all.
So I ended up using LLM + a tool which implements hard constraints and gives back validation data to LLM so the LLM can figure out why something wouldn't fit that specific way
Claude 4.6 Opus and Gemini 3.1 Pro can to some degree, although the 3D models they produce are often deficient in some way that my eval didn't capture.
My eval used OpenSCAD simply due to familiarity and not having time to experiment with build123d/CadQuery. There is an academic paper where they were successful at fine-tuning a small VLM to do CadQuery: https://arxiv.org/pdf/2505.14646
Great work - looks like building block towards 3d-model composition integration testing. I have been looking for a solution that would allow testing component fit into surrounding components. My use-case would be to create parametric boat hull and then add components to that could be tested for fitness in the arrangement.
Even with AI it is going to be incredibly difficult. Not really a full CAD kernel but I have been at something similar for the last 5 months - https://lilicad.com
I have a suggestion, a major part of 3D modelling that many can struggle with is actually UV mapping. It can take a bit of time to do to get it right and if it could be done decently with one click with better results than with Blender or paid applications like Rizom. They'd probably be a decent market out there.
CAD modeling seems to be safe from automation for the time being. I've tried various services and ones like sloyd.ai can't even take a simple svg and plop it onto a rectangle base.
And here I thought the CS dept in my school were the elite ones since they brought in the most money and sponsorships. Turns out my fellow Mech Eng classmates will have the last laugh.
My guess is there there is no internal pathways between the code representation of a CAD model and the language, concepts and experience of an object in 3D. They can often communicate to you the deficiencies of a picture of the model, but still fail to correct it.
Having played around with this a bit, I recommend using https://github.com/CadQuery/cadquery as the CAD language instead. I'm pretty sure it could even transpile to OnShape / Solidworks models as well, though it might require some funky hacks with their extensions frameworks
csgrs author here. Congratulations on this release! Please swing by the csgrs discord or let me know where to join you for future discussions related to SynapsCAD. I'd love to chat about future developments. Amazing work!
I use openscad vibe coding quite a bit. It tends to fall down beyond extremely simple examples, though. Error categories I've encountered if you want to build better rails: (1) forgetting axis orientations after multiple layers of rotate() (2) center=true presence ignored resulting in mistaken geometry (3) inconsistent naming (4) insufficiently verbose naming (5) 3D printing tolerances (6) lack of validation (7) shared faces causing rendering issues
I would suggest that every stage has the following basic checks: (A) If it's a 'substract' type operation, ensure the resulting shape has less volume than the original shape (B) Ensure no 'subtract' results in zero volume shape (C) Ensure no 'shared faces' exist (D) Ensure output is consistent with requisite axes (eg. render an elevation in orthographic and know which way is up/down in profile so that relative terms can be quantitatively verified in the rendering) (E) Name everything with a semantic tree that is updated properly instead of hacked upon until it becomes illogical and incoherent
This would go a huge way to fixing the main issues encountered so far.
If it’s GPL then you must care about “free software”. If that is the case, you should reconsider hosting on a fully-proprietary code forge. It requires cognitive dissonance to thing FLOSS is the right license for your project but not for your tooling/community.
You’re projecting that you would feel cognitive dissonance if you made the same choices as OP.
It’s perfectly reasonable for people to want to license their own work under a particular license without believing that everyone else should do the same, for example. Someone with that opinion, or a whole host of other positions that allow for different licenses to coexist, wouldn’t experience any cognitive dissonance.
The question then becomes what you are going to do about your cognitive dissonance. Continue to believe that “everyone else must be wrong”?
e.g. https://wiki.roshangeorge.dev/w/Blog/2026-01-11/Modeling_Wit...
Presumably someone is getting closer to this, curious who the most robust player in that space is.
Also curious if building an actual kernel replacement for open cascade is on the table now with AI, it’s a very tough thing to do but now it seems somewhat tractable in 2026
So I ended up using LLM + a tool which implements hard constraints and gives back validation data to LLM so the LLM can figure out why something wouldn't fit that specific way
Claude 4.6 Opus and Gemini 3.1 Pro can to some degree, although the 3D models they produce are often deficient in some way that my eval didn't capture.
My eval used OpenSCAD simply due to familiarity and not having time to experiment with build123d/CadQuery. There is an academic paper where they were successful at fine-tuning a small VLM to do CadQuery: https://arxiv.org/pdf/2505.14646
And here I thought the CS dept in my school were the elite ones since they brought in the most money and sponsorships. Turns out my fellow Mech Eng classmates will have the last laugh.
I would suggest that every stage has the following basic checks: (A) If it's a 'substract' type operation, ensure the resulting shape has less volume than the original shape (B) Ensure no 'subtract' results in zero volume shape (C) Ensure no 'shared faces' exist (D) Ensure output is consistent with requisite axes (eg. render an elevation in orthographic and know which way is up/down in profile so that relative terms can be quantitatively verified in the rendering) (E) Name everything with a semantic tree that is updated properly instead of hacked upon until it becomes illogical and incoherent
This would go a huge way to fixing the main issues encountered so far.
It’s perfectly reasonable for people to want to license their own work under a particular license without believing that everyone else should do the same, for example. Someone with that opinion, or a whole host of other positions that allow for different licenses to coexist, wouldn’t experience any cognitive dissonance.
The question then becomes what you are going to do about your cognitive dissonance. Continue to believe that “everyone else must be wrong”?